X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/863adae6473cc1fe177d0c48bc07b752beb83880..6b386cb2aadbfae5a94a139165dca0576fda7a46:/src/smpi/smpi_base.c diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c index dba43c97c7..76e98d8406 100644 --- a/src/smpi/smpi_base.c +++ b/src/smpi/smpi_base.c @@ -7,6 +7,7 @@ #include "private.h" #include "xbt/virtu.h" #include "mc/mc.h" +#include "mc/mc_replay.h" #include "xbt/replay.h" #include #include "simix/smx_private.h" @@ -195,7 +196,12 @@ double smpi_mpi_wtime(){ if (smpi_process_initialized() && !smpi_process_finalized() && !smpi_process_get_sampling()) { smpi_bench_end(); time = SIMIX_get_clock(); - //to avoid deadlocks if called too many times + // to avoid deadlocks if used as a break condition, such as + // while (MPI_Wtime(...) < time_limit) { + // .... + // } + // because the time will not normally advance when only calls to MPI_Wtime + // are made -> deadlock (MPI_Wtime never reaches the time limit) if(smpi_wtime_sleep > 0) simcall_process_sleep(smpi_wtime_sleep); smpi_bench_begin(); } else { @@ -250,10 +256,8 @@ static MPI_Request build_request(void *buf, int count, else request->refcount = 0; request->op = MPI_REPLACE; -#ifdef HAVE_TRACING request->send = 0; request->recv = 0; -#endif if (flags & SEND) smpi_datatype_unuse(datatype); return request; @@ -344,11 +348,14 @@ void smpi_mpi_start(MPI_Request request) if (request->flags & RECV) { print_request("New recv", request); + + xbt_mutex_t mut=smpi_process_mailboxes_mutex(); + xbt_mutex_acquire(mut); if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")){ //We have to check both mailboxes (because SSEND messages are sent to the large mbox). begin with the more appropriate one : the small one. mailbox = smpi_process_mailbox_small(); - XBT_DEBUG("Is there a corresponding send already posted the small mailbox %p (in case of SSEND)?", mailbox); + XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox); smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request); if(action ==NULL){ @@ -356,7 +363,7 @@ void smpi_mpi_start(MPI_Request request) XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox); action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request); if(action ==NULL){ - XBT_DEBUG("Still notching, switch back to the small mailbox : %p", mailbox); + XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox); mailbox = smpi_process_mailbox_small(); } }else{ @@ -393,18 +400,16 @@ void smpi_mpi_start(MPI_Request request) request, -1.0); XBT_DEBUG("recv simcall posted"); - + xbt_mutex_release(mut); } else { int receiver = request->dst; - #ifdef HAVE_TRACING - int rank = request->src; - if (TRACE_smpi_view_internals()) { - TRACE_smpi_send(rank, rank, receiver,request->size); - } - #endif + int rank = request->src; + if (TRACE_smpi_view_internals()) { + TRACE_smpi_send(rank, rank, receiver,request->size); + } print_request("New send", request); //if we are giving back the control to the user without waiting for completion, we have to inject timings @@ -419,6 +424,9 @@ void smpi_mpi_start(MPI_Request request) XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size)); } + xbt_mutex_t mut=smpi_process_remote_mailboxes_mutex(receiver); + xbt_mutex_acquire(mut); + if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode mailbox = smpi_process_remote_mailbox(receiver); XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox); @@ -454,11 +462,11 @@ void smpi_mpi_start(MPI_Request request) oldbuf = request->buf; if (!smpi_process_get_replaying() && oldbuf && request->size!=0){ if((smpi_privatize_global_variables) - && ((char*)request->buf >= start_data_exe) - && ((char*)request->buf < start_data_exe + size_data_exe )){ + && ((char*) request->buf >= smpi_start_data_exe) + && ((char*)request->buf < smpi_start_data_exe + smpi_size_data_exe )){ XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment "); - smpi_switch_data_segment(request->src); - } + smpi_switch_data_segment(request->src); + } buf = xbt_malloc(request->size); memcpy(buf,oldbuf,request->size); XBT_DEBUG("buf %p copied into %p",oldbuf,buf); @@ -484,13 +492,11 @@ void smpi_mpi_start(MPI_Request request) -#ifdef HAVE_TRACING /* FIXME: detached sends are not traceable (request->action == NULL) */ if (request->action) - simcall_set_category(request->action, TRACE_internal_smpi_get_category()); - -#endif + simcall_set_category(request->action, TRACE_internal_smpi_get_category()); + xbt_mutex_release(mut); } } @@ -694,8 +700,8 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) if((req->flags & ACCUMULATE) || (datatype->has_subtype == 1)){ if (!smpi_process_get_replaying()){ if( smpi_privatize_global_variables - && ((char*)req->old_buf >= start_data_exe) - && ((char*)req->old_buf < start_data_exe + size_data_exe ) + && ((char*)req->old_buf >= smpi_start_data_exe) + && ((char*)req->old_buf < smpi_start_data_exe + smpi_size_data_exe ) ){ XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment "); smpi_switch_data_segment(smpi_process_index()); @@ -719,7 +725,6 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) } -#ifdef HAVE_TRACING if (TRACE_smpi_view_internals()) { if(req->flags & RECV){ int rank = smpi_process_index(); @@ -727,7 +732,6 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) TRACE_smpi_recv(rank, src_traced, rank); } } -#endif if(req->detached_sender!=NULL){ smpi_mpi_request_free(&(req->detached_sender)); @@ -745,8 +749,12 @@ int smpi_mpi_test(MPI_Request * request, MPI_Status * status) { //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before) - //to avoid deadlocks - //multiplier to the sleeptime, to increase speed of execution, each failed test will increase it + // to avoid deadlocks if used as a break condition, such as + // while (MPI_Test(request, flag, status) && flag) { + // } + // because the time will not normally advance when only calls to MPI_Test + // are made -> deadlock + // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it static int nsleeps = 1; if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep); @@ -855,8 +863,9 @@ void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* MPI_Request request =build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag, comm, PERSISTENT | RECV); - //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls - //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it + // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls + // (especially when used as a break condition, such as while(MPI_Iprobe(...)) ... ) + // multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it static int nsleeps = 1; if(smpi_iprobe_sleep > 0) simcall_process_sleep(nsleeps*smpi_iprobe_sleep); // behave like a receive, but don't do it